Deployment and Fine-Tuning Issues with Phi-4 on Serverless API

[mrzaizai2k]

1. Missing Task Type Selection
   Unlike the Phi-4 Mini model, the Phi-4 fine-tuning interface does not provide an option to select the task type

This is when finetuning with Phi 4

This is when finetuning with Phi 4 Mini

2. Serverless API Option Not Available
   When accessing the Phi-4 model via the Fine-Tune tab without applying catalog filters, the Serverless API option is not visible. But in Model Catalog tab, it said we can finetune via Serverless. This inconsistency may cause confusion for users expecting uniform access across different navigation paths.

When using Model Catalog tab

I check in detail, and we don't have any problem with region (Eastus2) and quota, I have hub-project based. I check all the documents and see no related issues

When using Fine tuning tab

[mrzaizai2k]

I can deploy and finetune Phi 4 Mini with the same steps, and it worked. Don't know what happened with Phi 4. Thank you

[amynic]

adding @kinfey and @nitya from my team who work with Phi4 and Finetuning respectively - can we explore the issue above? @kinfey looks like the focus is fine tuning with Phi4 in Foundry

[mrzaizai2k]

Yes, I can just fine-tune phi 4 mini

…

On Mon, Aug 4, 2025 at 5:35 PM Kinfey ***@***.***> wrote: I think just Phi-4-mini can be fine tuning in Azure AI Foundry — Reply to this email directly, view it on GitHub <#119 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJYP3L42F2CFUPQ4BSMGBTD3L4ZNXAVCNFSM6AAAAACDA4DTHKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGOJYGY4TSNQ> . You are receiving this because you authored the thread.Message ID: <azure-ai-foundry/discussions/repo-discussions/119/comments/13986996@ github.com>

[kinfey]

@mrzaizai2k use hub resource , I think it can ft

[mrzaizai2k]

Hmm That's weird. As this link fine-tuning-overview

Fine-Tuning Availability
Now that you know when to use fine-tuning for your use case, you can go to Azure AI Foundry to find models available to fine-tune.

To fine-tune an AI Foundry model using Serverless you must have a hub/project in the region where the model is available for fine tuning. See Region availability for models in serverless API deployment for detailed information on model and region availability, and How to Create a Hub based project to create your project.

And I have a hub based project as this

However, I will try to create a hub and see if it can fix this, I will give the answer later

[mrzaizai2k]

Got it — you're using Managed Compute. What I'm aiming to do is fine-tune Phi-4 using the Serverless API. As I mentioned earlier, when I filter by Model Catalog, it indicates that Phi-4 is available for fine-tuning via the Serverless API (as shown in the image above). However, when I switch to the Fine-tuning tab, Phi-4 doesn't appear as a Serverless option. That's what's causing the confusion.

[mrzaizai2k]

This is new, I now can deploy Phi 4 on Chrome browser, but still can not finetune it
I change the browser after this answer: https://learn.microsoft.com/en-us/answers/questions/5513052/deployment-and-fine-tuning-issues-with-phi-4-on-se?page=1&orderby=newest&comment=answer-12169727&translated=false#newest-answer-comment

[mrzaizai2k]

But I still need to finetune Phi 4, I still have problem with task type. pls

[nitya]

@mrzaizai2k - I was able to reproduce your issue and will reach out to the internal team to see how best to respond. Note that Phi-4 Serverless Fine-tuning is in preview so your feedback is really appreciated!!

In the meantime I have one suggestion for you.

At the top of each Azure AI Foundry portal page you will see a smiley face next to the settings gear icon (see screenshot below) - clicking that will open the "Send us feedback" dialog as shown below. This is the fastest way to directly report issues with the portal experience.

Fill in your experience there and reference this discussion Item for the screenshots you provided (since dialogs are harder to capture with the auto-screenshot experience) and check the box if you want the Microsoft team to contact you directly for more clarification.

[mrzaizai2k]

New thing. When I click "Fine-tune" for the first time, the UI displays the "Compute" tab, allowing me to select a machine (Managed Compute).

However, if I cancel that action and click "Fine-tune" again, the "Compute" tab no longer appears. Instead, it defaults to Serverless, and I don’t see any option to choose between Managed Compute and Serverless.

There’s no visible button or setting that lets me switch the compute type after the first attempt. I believe this is a UI bug.

[nitya]

I was able to validate this flow as well. Thank you so much for the detailed reporting - appreciate it!

1. Selected Phi-4 from model catalog
2. Picked Fine tune option in model card => gave me managed option
3. Cancelled it => got back to model card view
4. Picked Fine tune option in model card => gave me serverless option

After this, any further retries in model card view => always give serverless
If I return to catalog and start again => gets back to step 2 (managed) then repeats 3-4.

Since the Fine-tune Phi-4 experience is in preview, this is useful feedback. I will reach out internally to share this as well - but please do report it via the "smiley icon" at the top of the Azure AI Portal so the team gets more signals about this.

[mrzaizai2k]

Thank you